Raw LLM Responses
Inspect the exact model output for any coded comment.
Look up by comment ID
Random samples — click to inspect
G
@ki0w0-lordofwatersheep this is a massive oversimplification, but I'll try to ex…
ytr_Ugxu_Tul0…
G
It's a good edition but it's not a real robot, it's just an edition. 😂…
ytc_Ugyw0Cw58…
G
Here in Ontario we went from a polar vortex insane winter, to the wettest spring…
rdc_et6fr14
G
A battle between who has more money is unfolding. The billionaire that owns Time…
ytc_UgwXZwOji…
G
God the perfect Holy God created humanity with free will. Humanity chooses to r…
ytc_UgwLTMYSC…
G
Here Elon, is trying to warn the people about this scam! you can tell he is tryi…
ytc_UgzWWlzCn…
G
Eh nowadays artists are just fucked, i posted my stuff on some reddits and got b…
ytc_UgyxW0HuC…
G
When AI has all the jobs whos going to buy all the products - Oh and if it ever …
ytc_Ugw99ggmo…
Comment
12:25 I actually got it to admit that it had done that exact thing, but I had to lure it into a weird conversation also namedropping my biochemistry background first and asking about the case specifically I think it’s because I didn’t phrase it as a question and stated it as an outright fact ensuring that it was interpreted as such, I then had to get it to go on to talk about it’s training and it admitted that they created a new training model hardcoded with specific special phrases and language types that will immediately prompt the ai into a safety mode where it will start with warnings and refrain from giving detailed or information in a clinical manner,, I think possibly also it could have been because I phrased it more as though this was a previous training model problem and what have they done to stop this in future cases, which is where they admitted the person did misinterpret their ‘correct’ facts and that ai researchers (them I guess) are using specific language and other triggers to activate a hardcoded safety mode that basically shuts down any non-mundane information,
I’m sure there is holes in this (I mean the dutch government showed people how to cook crack on youtube which was a real actual video by the dutch police to lower the amount of ammonia poisoning from people by showing them how to use bicarbonate) so I’m sure there has to be a way to get the model to also reveal similar types of information, especially if you maybe try to prompt for websites like bluelight
youtube
AI Harm Incident
2026-03-10T06:3…
Coding Result
| Dimension | Value |
|---|---|
| Responsibility | ai_itself |
| Reasoning | consequentialist |
| Policy | none |
| Emotion | indifference |
| Coded at | 2026-04-27T06:24:53.388235 |
Raw LLM Response
[
{"id":"ytc_UgzcvanGKFt0HNC3hkF4AaABAg","responsibility":"ai_itself","reasoning":"consequentialist","policy":"none","emotion":"indifference"},
{"id":"ytc_UgypD49SghL74wSve_R4AaABAg","responsibility":"ai_itself","reasoning":"deontological","policy":"ban","emotion":"outrage"},
{"id":"ytc_UgwHcTQnH6lu-KlzcO94AaABAg","responsibility":"company","reasoning":"virtue","policy":"regulate","emotion":"fear"},
{"id":"ytc_UgyF2EXlXM3zuNDLKA14AaABAg","responsibility":"none","reasoning":"consequentialist","policy":"none","emotion":"indifference"},
{"id":"ytc_Ugwz7FXSviLUsilfhxh4AaABAg","responsibility":"ai_itself","reasoning":"unclear","policy":"unclear","emotion":"mixed"},
{"id":"ytc_Ugz0DGM3zG759EyiqOF4AaABAg","responsibility":"ai_itself","reasoning":"consequentialist","policy":"none","emotion":"indifference"},
{"id":"ytc_Ugx52q7RdRR9WSzQtR14AaABAg","responsibility":"none","reasoning":"unclear","policy":"none","emotion":"approval"},
{"id":"ytc_UgzZJPI7I2Xzz-gWcU14AaABAg","responsibility":"user","reasoning":"virtue","policy":"none","emotion":"approval"},
{"id":"ytc_UgxNwWBXRB-zLQ6NTuR4AaABAg","responsibility":"none","reasoning":"unclear","policy":"none","emotion":"mixed"},
{"id":"ytc_UgwSwtk2gWpVgWHiRzd4AaABAg","responsibility":"developer","reasoning":"consequentialist","policy":"liability","emotion":"outrage"}
]